DeepSeek Harness Status: rc.7, MIT, Developer Preview
The short answer: DeepSeek Harness is an official open-source project by DeepSeek AI (MIT), currently in Developer Preview. Baseline:
0.1.0-rc.7(commit99f6f02, 2026-08-17).npx @deepseek-ai/dsh webruns it with zero installation. The official README warns plainly: "THERE WILL BE COMPATIBILITY-BREAKING CHANGES." Using it means pinning versions and reading changelogs.
This is a status page — all facts are cross-checked against the official repository
deepseek-ai/deepseek-harness, and the baseline moves with each release. Written 2026-08-17.
Status overview
| Item | Value |
|---|---|
| Repository | deepseek-ai/deepseek-harness (official org) |
| Version | 0.1.0-rc.7 (npm package @deepseek-ai/dsh, @next tag; latest stays rc.6) |
| Audit baseline | commit 99f6f02 (2026-08-17) |
| License | MIT (third-party disclosures in THIRD_PARTY_NOTICES.md) |
| Status | Developer Preview, breaking changes expected (stated in the README) |
| Node requirement | ^22.19.0 || >=24.0.0 (root package.json#engines) |
| Install | npx @deepseek-ai/dsh web (zero install); from source see Quick Start |
| Web UI | http://127.0.0.1:3080 by default (loopback) |
What works today
Per the official docs/user/guide/, DSH already covers a usable agent workbench:
| Capability | Official doc |
|---|---|
| Web UI workbench: model config, workspace selection, task execution, approval flow | docs/user/guide/index.md |
| Multi-model and custom OpenAI-compatible endpoints | docs/user/guide/providers.md |
| Python SDK (drive a Harness from another process) | docs/user/guide/python-sdk.md |
| Other CLI modes (headless one-shot tasks, etc.) | apps/cli/README.md |
| Plugin development | docs/develop/ |
| Architecture and development guides | docs/architecture.md, docs/development.md |
The official docs ship Chinese versions too (index.zh.md, providers.zh.md, python-sdk.zh.md). This site adds what the official docs don't: deep dives — Capabilities, Core mechanics, Plugin development.
Compatibility risks and countermeasures
Developer Preview is not a slogan; the README says it outright: "THERE WILL BE COMPATIBILITY-BREAKING CHANGES." Countermeasures:
- Pin versions: exact npm version (
npx @deepseek-ai/dsh@0.1.0-rc.7) or a lockfile; when running from source,git checkouta known baseline - Read changelogs before upgrading: official releases and commit history
- Verify the composition tree: after an upgrade, run
dsh web --dump-configto confirm the plugin assembly (Boot config) - Plugin compatibility: third-party plugins can lag core changes — upgrade plugins together with the core, don't mix versions
How to follow along
| Channel | Purpose |
|---|---|
| GitHub Discussions | Feedback and Q&A (official README) |
| Discord | Real-time community (official README) |
GitHub Releases / npm @deepseek-ai/dsh | Release cadence |
dsh-plugin topic | Plugin ecosystem discovery (official recommendation for plugin repos) |
| This site's Plugin ecosystem | Community-maintained selection |
Frequently asked questions
Has DeepSeek Harness been officially released?
The repository is public and the npm package runs with zero install — but the version is 0.1.0-rc.7: Developer Preview, not stable. Getting work done with it is fine; treating it as a stable API will hurt.
Is DSH official from DeepSeek?
Yes. It lives in the deepseek-ai org and the README credits DeepSeek AI. This site is a community tutorial site with no affiliation (What is DSH).
Does DeepSeek publish Chinese documentation?
Yes. docs/user/guide/ carries index.zh.md, providers.zh.md, python-sdk.zh.md alongside the English versions. This site complements, not copies: deep dives (mechanics, plugin development, troubleshooting).
Can it be used in production?
The official status is Developer Preview, but the privacy baseline is solid: loopback-only listener, 0600-managed credentials, restricted sandbox by default (Sandbox & security). Decision points: pin versions, read changelogs, vet the plugin supply chain. The full production trade-off is in the comparison page FAQ.
Conclusion
DSH is at the stage of "usable, iterating fast, officially admitting it breaks compatibility." The reasons to choose it (open source, model freedom, local) don't change; the discipline it demands is version pinning. This baseline will be updated at the next audit.